A Deep Learning-Assisted Method for Camellia Oleifera Trunk Detection Using the Enhanced Yolov8-SEAW Model

Abstract

Efficient detection of Camellia oleifera trunks in complex environments is vital for advancing intelligent harvesting robotics. However, challenges such as occlusion, background noise, and varying trunk shapes often lead to missed detections and false positives. To address these, this study presents YOLOv8-SEAW, an enhanced version of the YOLOv8 model designed to improve detection accuracy in such conditions. The architecture integrates three key improvements: SPD-Conv boosts small target detection, Efficiency RepGFPN enhances multiscale feature fusion, and ACmix attention minimizes background interference. The WIoUv1 loss function further refines bounding box regression, improving localization accuracy. Experimental results show that YOLOv8-SEAW boosts mAP from 82.4% to 89.4% and precision from 70% to 96%, with a 26% relative increase. Additionally, the model reduces parameters from 3.1 million to 2.9 million, improving efficiency without sacrificing accuracy. Overall, YOLOv8-SEAW enhances trunk detection, particularly in cluttered and occluded scenes, and is well-suited for deployment in automated agricultural tasks.